Nice framing of the computing and UX issues. IMHO, Intercom’s UX strikes a better balance between deterministic structure and probabilistic flexibility—critical for scaling complex workflows.
Totally agree.. that's the question isn't it. We actually are building something very similar here as our user base is also more used to Documents as the output of their work so it feels natural that the input and output of an Agent is also a document
What we have not nailed, collectively, is showing the difference between a deterministic decision and probabilistic judgment. The input UI often looks the same for both, but the output semantics are not. A Yes No answer (deterministic) has a single path. A Maybe (non-deterministic) is a distribution, think of a confusion matrix from machine learning.
I would love to see agent UIs that surface confidence, failure modes, and alternatives by default. Kinda like a micro confusion-matrix UI and a first class “defer or ask” state when the model is unsure.
Great read. Agree entirely that we still haven’t found the right set of primitives for blending deterministic and probabilistic systems.
A product I’ve been really enjoying using is Klaro (https://www.klaroapp.com/) which feels like an interesting fusion of a canvas, notebook and AI-powered document to help users work and “play” with numbers.
What an amazing read - thank you sharing. Going to feature this on aiverse [dot] design
good read - amazed by what was the thought process behind building fin.
So good. Thank you for taking the time to break this all down like this.
Nice framing of the computing and UX issues. IMHO, Intercom’s UX strikes a better balance between deterministic structure and probabilistic flexibility—critical for scaling complex workflows.
Great post thank you - puts words to the struggle we’ve had the last year and validated our current approach!
Totally agree.. that's the question isn't it. We actually are building something very similar here as our user base is also more used to Documents as the output of their work so it feels natural that the input and output of an Agent is also a document
What we have not nailed, collectively, is showing the difference between a deterministic decision and probabilistic judgment. The input UI often looks the same for both, but the output semantics are not. A Yes No answer (deterministic) has a single path. A Maybe (non-deterministic) is a distribution, think of a confusion matrix from machine learning.
I would love to see agent UIs that surface confidence, failure modes, and alternatives by default. Kinda like a micro confusion-matrix UI and a first class “defer or ask” state when the model is unsure.
Its fun time to design UI and interaction
Really happy I found this because that was super interesting to read. Thanks for sharing. Our team is super happy with Fin so keep up the great work!
Great to hear, cheers Daniel
Great read. Agree entirely that we still haven’t found the right set of primitives for blending deterministic and probabilistic systems.
A product I’ve been really enjoying using is Klaro (https://www.klaroapp.com/) which feels like an interesting fusion of a canvas, notebook and AI-powered document to help users work and “play” with numbers.
Yeah, definitely going to be interesting to see how these patterns evolve. Interesting times ahead!